A material attribute intelligent dialogue method and system based on knowledge graph embedding and diffusion generation mechanism

Abstract

The application discloses a material attribute intelligent dialogue method and system based on a knowledge graph embedding and diffusion generation mechanism, and belongs to the technical field of the cross of artificial intelligence and material science. The application solves the problems that existing methods are faced with data scarcity, insufficient prediction ability for complex crystal structures and inability to perform data privacy protection. The application firstly constructs material knowledge graph data through entity extraction; secondly, a graph neural network conditional diffusion model is combined to construct a crystal structure containing alloy components; then, a self-supervised independent pre-training model is used for general feature extraction; finally, semantic features, crystal structure features, entity features and knowledge graph features are fused, and the fused features are processed to serve as input of a generative language model to generate professional dialogue responses. The method can be applied to material attribute intelligent dialogue.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and materials science, specifically relating to a method and system for intelligent dialogue on material properties based on knowledge graph embedding and diffusion generation mechanisms. Background Technology

[0002] Magnesium alloys have high application value in aerospace and transportation fields due to their low density and excellent mechanical properties. However, due to their unique multi-principal element design concept, the calculation of thermodynamic parameters such as mixing entropy and enthalpy within the alloy system is extremely complex. Traditional trial-and-error methods and basic alloy theory are difficult to efficiently and accurately guide their composition design and performance prediction.

[0003] Although machine learning technology has begun to be applied to materials research and development, it still faces challenges such as data scarcity and insufficient predictive ability of models for complex crystal structures. In addition, the requirements for privacy protection of materials data are becoming increasingly stringent.

[0004] Traditional centralized machine learning training requires aggregating experimental data from various institutions to a central server, which poses significant risks of data leakage and compliance issues, especially when dealing with high-value alloy formulations. Therefore, there is an urgent need for an intelligent method that can guide magnesium alloy composition design and performance prediction while meeting the privacy protection requirements of materials data. Summary of the Invention

[0005] The purpose of this invention is to address the problems of data scarcity, insufficient predictive ability for complex crystal structures, and inability to protect data privacy in existing methods. Therefore, this invention proposes a material property intelligent dialogue method and system based on knowledge graph embedding and diffusion generation mechanism.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] A material property intelligent dialogue method based on knowledge graph embedding and diffusion generation mechanism, the method specifically includes the following steps:

[0008] Step 1: Obtain the labeled material dataset;

[0009] Step 2: Construct a knowledge graph based on the acquired material dataset. After generating the knowledge graph, construct node, edge, and community indexes for the knowledge graph data.

[0010] Step 3: Perform semantic decomposition on the dialogue context of the current query, extract the target attributes of magnesium alloy expected by the user, and retrieve the constructed knowledge graph based on the extracted target attributes of magnesium alloy expected by the user to retrieve the entity nodes and relation edges related to the current query.

[0011] Step 4: Convert the retrieved entity nodes and relation edges into graph historical feature vectors. The semantic feature vector of the current query dialogue context is extracted using a text encoder. ;

[0012] Construct a diffusion model based on U-Net, and replace the convolutional layers in U-Net with a multi-layer graph convolutional network and a graph attention network;

[0013] The user's desired magnesium alloy target properties and noise are used as inputs to the diffusion model, which generates a crystal structure containing the alloy composition.

[0014] Step 5: Use a crystal structure encoder to extract features from the physical crystal structure generated in Step 4 to obtain the crystal structure feature vector. ;

[0015] Step 6: Identify entities in the current query context through an entity prediction mechanism and extract feature vectors of the identified entities. Then , , and After splicing, the splicing result is input into the masking multi-cross attention mechanism layer;

[0016] Step 7: The output of the masked multi-cross attention mechanism layer is weighted and fused by average pooling operation. The fusion result is fed into the generative language model. The generative language model uses the temporal mechanism of right-shifting output to generate a response containing graph reasoning, crystal structure and material properties.

[0017] Furthermore, the knowledge graph is constructed as follows:

[0018] The BERT-BiLSTM-Att-CRF model was used to label the material data in the training set in BIO format. Then, the GraphRAG retrieval tool was used to recall entities and relationships in the BIO-labeled dataset. The entities include magnesium alloy composition, crystal structure parameters and mechanical property indicators.

[0019] The recalled entities and relationships are then format-converted and aggregated. The aggregation results are then converted into adjacency lists and natural language forms to input the generator. The generator output is then converted into code snippets and syntax trees. Finally, the code snippets and syntax trees are represented as node sequences and graph embeddings. The node sequences and graph embeddings are then aggregated by the graph generator to generate a knowledge graph containing nodes and edges.

[0020] Furthermore, the knowledge graph is constructed as follows:

[0021] We use the DeepSeek R1 generative model to generate natural text responses with hidden entities for dialogues in the dataset, and extract the sparse attention weights hidden in the natural text responses. Then, we use the alignment parsing module based on rules and semantic similarity to map the sparse attention weights into structured entities and relationships. Finally, we import the parsed entities and relationships into GraphRAG to construct a dynamically evolving knowledge graph.

[0022] Furthermore, the step of using the user's desired magnesium alloy target properties and noise as input to the diffusion model, and generating the crystal structure through the diffusion model, specifically involves:

[0023] Step 4.1 Initialize the time step The complete noise data and time-step encoding are used as input to the constructed diffusion model;

[0024] Step 42: Extract node features (i.e., component features containing atomic type information) and topological edge features through multi-layer graph convolutional networks or graph attention networks within the diffusion model;

[0025] Step 4.3: Perform cross-attention calculation on the target attribute and the extracted node features. That is, use the target attribute as the query vector, the extracted node features as the key vector and value vector, and use the cross-attention calculation result as the updated node features to continue propagating forward until the denoised crystal structure containing the alloy composition is output at the current time step.

[0026] Step 44: Determine if the condition is met. :

[0027] If satisfied, then execute steps four and five on the crystal structure output at the last time step;

[0028] If not satisfied, then let , will the The denoised crystal structure and time-step encoding after each time step are used as input to the constructed diffusion model, and the process returns to step four two.

[0029] Steps four and five: Based on the physical conservation laws and spatial topology rules of materials science, the types of atoms (to determine the final alloy composition), atomic positions, and lattice parameters in the crystal structure output in the last time step are corrected and discretized and mapped, and finally the physical crystal structure corresponding to the current query context dialogue is output.

[0030] Furthermore, the training process for the text encoder and the crystal structure encoder is as follows:

[0031] Step 1: Dataset Collection and Text Preprocessing: Collect multimodal datasets containing magnesium alloy composition, crystal structure, and mechanical properties, and perform cleaning, encoding, sequence padding, and normalization on the collected datasets;

[0032] Step 2, Self-supervised pre-training:

[0033] For text data: Set up masked language modeling and autoregressive language modeling tasks, and pre-train the text encoder on the set tasks;

[0034] For crystal structure data: Set up mask atom prediction and structure denoising and reconstruction tasks, and pre-train the crystal structure encoder on the set tasks.

[0035] A material property intelligent dialogue system based on knowledge graph embedding and diffusion generation mechanism, the system includes a knowledge graph construction module, a self-supervised pre-training module, a diffusion model, and a generative language model, wherein:

[0036] The knowledge graph construction module is used to generate a knowledge graph based on the materials dataset.

[0037] The self-supervised pre-training module is used to train the text encoder and the crystal structure encoder.

[0038] The diffusion model is a U-Net-based diffusion model, in which the convolutional layers in U-Net are replaced by multi-layer graph convolutional networks or graph attention networks. The diffusion model takes the complete noisy data and the target attributes of the current dialogue context as input to generate a crystal structure containing alloy components.

[0039] The generative language model is used to generate responses that include graph reasoning, crystal structure, and material properties.

[0040] Furthermore, the knowledge graph construction method is as follows:

[0041] The BERT-BiLSTM-Att-CRF model was used to label the material data in the training set in BIO format. Then, the GraphRAG retrieval tool was used to recall entities and relationships in the BIO-labeled dataset. The entities include magnesium alloy composition, crystal structure parameters and mechanical property indicators.

[0042] The recalled entities and relationships are then format-converted and aggregated. The aggregation results are then converted into adjacency lists and natural language forms to input the generator. The generator output is then converted into code snippets and syntax trees. Finally, the code snippets and syntax trees are represented as node sequences and graph embeddings. The node sequences and graph embeddings are then aggregated by the graph generator to generate a knowledge graph containing nodes and edges.

[0043] Furthermore, the knowledge graph construction method is as follows:

[0044] We use the DeepSeek R1 generative model to generate natural text responses with hidden entities for dialogues in the dataset, and extract the sparse attention weights hidden in the natural text responses. Then, we use the alignment parsing module based on rules and semantic similarity to map the sparse attention weights into structured entities and relationships. Finally, we import the parsed entities and relationships into GraphRAG to construct a dynamically evolving knowledge graph.

[0045] Furthermore, the working process of the self-supervised pre-training module is as follows:

[0046] Step 1: Dataset Collection and Text Preprocessing: Collect multimodal datasets containing magnesium alloy composition, crystal structure, and mechanical properties, and perform cleaning, encoding, sequence padding, and normalization on the collected datasets;

[0047] Step 2: For text data: Set up masked language modeling and autoregressive language modeling tasks, and pre-train the text encoder on the set tasks;

[0048] For crystal structure data: Set up mask atom prediction and structure denoising and reconstruction tasks, and pre-train the crystal structure encoder on the set tasks.

[0049] Furthermore, the working process of the diffusion model is as follows:

[0050] Step A: Initialize the time step The complete noise data and time-step encoding are used as input to the constructed diffusion model;

[0051] Step B: Extract node features and topological edge features through a multi-layer graph convolutional network or graph attention network within the diffusion model;

[0052] Step C: Perform cross-attention calculation on the target attribute and the extracted node features. That is, use the target attribute as the query vector, the extracted node features as the key vector and value vector, and use the cross-attention calculation result as the updated node features to continue propagating forward until the denoised crystal structure containing alloy composition is output at the current time step.

[0053] Step D: Determine if the condition is met. :

[0054] If satisfied, then execute step E on the crystal structure containing alloy composition output from the last time step;

[0055] If not satisfied, then let , will the The denoised crystal structure and time-step encoding after each time step are used as input to the constructed diffusion model, and the process returns to step B.

[0056] Step E: Based on the physical conservation laws and spatial topology rules of materials science, the atomic types, atomic positions, and lattice parameters in the crystal structure output at the last time step are corrected and discretized, and finally the physical crystal structure corresponding to the current query context dialogue is output.

[0057] The beneficial effects of this invention are:

[0058] This invention combines a graph neural network conditional diffusion model to reverse-engineer and generate physically reasonable alloy compositions and crystal structures, improving the predictive ability for complex crystal structures and solving the problem that traditional models cannot handle the generation of complex materials, without requiring a large amount of training data. Through the deep integration of GraphRAG technology and a crystal structure encoder, this invention achieves multimodal accurate inference, enhancing the model's ability to infer deep material relationships and significantly improving the accuracy of R&D question answering. Furthermore, this invention balances privacy compliance with model evolution, employing a federated persistence and forgetting mechanism to flexibly withdraw or add data contributions without aggregating sensitive experimental data, meeting compliance requirements and ensuring dynamic maintenance. Attached Figure Description

[0059] Figure 1 This is a diagram illustrating the human-machine collaborative workflow and overall framework of the present invention.

[0060] Figure 2 This is a flowchart of the knowledge graph construction process in method (1) of the present invention;

[0061] Figure 3 This is a generation model diagram of the diffusion mechanism of the present invention;

[0062] Figure 4 This is a flowchart of the multimodal dialogue generation stage of the present invention;

[0063] Figure 5 This is a schematic diagram of the overall framework and retraining path of the federated continuous / forgetful learning of the present invention. Detailed Implementation

[0064] Specific implementation method one: Combining Figure 4 This embodiment describes a material property intelligent dialogue method based on knowledge graph embedding and diffusion generation mechanisms. The method specifically includes the following steps:

[0065] Step 1: Obtain the labeled material dataset. The material data includes material data with standard attributes and material data without standard attributes. The material data without standard attributes needs to be labeled manually.

[0066] Step 2: Construct a knowledge graph based on the acquired material dataset. After generating the knowledge graph, construct node, edge, and community indexes for the knowledge graph data to support efficient retrieval in subsequent steps.

[0067] Specifically, step two is achieved through either method (1) or method (2):

[0068] Method (1), such as Figure 2 As shown, the BERT-BiLSTM-Att-CRF model is used to label the material data in the training set in BIO format. Then, the GraphRAG retrieval tool (which includes query expansion, query decomposition, merging and pruning modules) is used to recall entities and relationships in the BIO format-labeled dataset. The entities include magnesium alloy composition, crystal structure parameters and mechanical property indicators.

[0069] The recalled entities and relationships are then format-converted and aggregated. The aggregation results are then converted into adjacency lists and natural language forms to input the generator. The generator output is then converted into code snippets and syntax trees (to enhance the logic). Finally, the code snippets and syntax trees are represented as node sequences and graph embeddings. The node sequences and graph embeddings are then aggregated by the graph generator to generate a knowledge graph containing nodes and edges.

[0070] Method (2): Use the DeepSeek R1 generative model to generate natural text responses with hidden entities for dialogues in the dataset, extract the sparse attention weights hidden in the natural text responses, and then use the Alignment & Parsing Module based on rules and semantic similarity to map the sparse attention weights into structured entities and relationships. Finally, import the parsed entities and relationships into GraphRAG to construct a dynamically evolving knowledge graph.

[0071] Step 3: Perform semantic decomposition on the dialogue context of the current query (analyze the user's R&D needs through the dialogue model), extract the target attributes of magnesium alloy expected by the user (such as specific yield strength), and retrieve the constructed knowledge graph based on the extracted target attributes of magnesium alloy expected by the user to retrieve the entity nodes and relation edges related to the current query.

[0072] Step 4: Convert the retrieved entity nodes and relation edges into graph historical feature vectors. The semantic feature vector of the current query dialogue context is extracted using a text encoder. ;

[0073] like Figure 3As shown, a diffusion model based on U-Net is constructed, and the convolutional layers in U-Net are replaced with multi-layer graph convolutional networks (GCN) or graph attention networks (GAT).

[0074] The forward diffusion process of the diffusion model of the present invention is as follows: starting with a clean crystal structure, Gaussian noise is gradually added to change the position information of atoms in the crystal structure, and to simulate the generation process of defects such as structural disturbance, atom shedding, and dislocation formation under temperature or pressure change conditions.

[0075] The desired magnesium alloy properties and noise are used as inputs to the diffusion model. The model employs a reverse diffusion process for gradual denoising, generating a crystal structure containing the alloy composition. Specifically:

[0076] Step 4.1 Initialize the time step The complete noise data and time-step encoding are used as input to the constructed diffusion model;

[0077] Step 42: Extract node features (e.g., compositional features of magnesium alloy atom type information) and topological edge features through multi-layer graph convolutional network (GCN) or graph attention network (GAT) in the diffusion model to complete the feature dimensionality reduction and encoding of the high-dimensional latent space;

[0078] Step 4.3: Perform cross-attention calculation on the target attribute and the extracted node features. That is, use the target attribute as the query vector, the extracted node features as the key vector and the value vector, and use the cross-attention calculation result as the updated node features to continue to propagate forward. The cross-attention guides the denoising process to focus on the local atomic structures that are strongly related to the target attribute until the denoised crystal structure containing the alloy composition at the current time step is output.

[0079] By utilizing the skip-connections mechanism of U-Net, the decoder fuses deep semantic features with shallow high-resolution structural features during the denoising process, and gradually predicts the distribution of noise added at the current time step.

[0080] Step 44: Determine if the condition is met. :

[0081] If satisfied, then execute steps four and five on the crystal structure containing alloy components output at the last time step;

[0082] If not satisfied, then let , will the The denoised crystal structure and time-step encoding after each time step are used as input to the constructed diffusion model, and the process returns to step four two.

[0083] Steps four and five: Based on the physical conservation laws and spatial topology rules of materials science, the types of atoms (to determine the final alloy composition), atomic positions, and lattice parameters in the crystal structure output in the last time step are corrected and discretized and mapped, and finally the physical crystal structure corresponding to the current query context is output.

[0084] Step 5: Use a crystal structure encoder to extract features from the physical crystal structure generated in Step 4 to obtain the crystal structure feature vector. ;

[0085] In this invention, the training process for the text encoder and the crystal structure encoder is as follows:

[0086] Step 1: Dataset Collection and Text Preprocessing: Collect multimodal datasets containing magnesium alloy composition, crystal structure, and mechanical properties, and perform cleaning, encoding, sequence padding, and normalization on the collected datasets;

[0087] Step 2, Self-supervised pre-training:

[0088] For text data: Set up masked language modeling (MLM) and autoregressive language modeling (ARLM) tasks, and pre-train the text encoder on the set tasks. The trained text encoder can extract domain semantic features.

[0089] For crystal structure data: Set up masked atom prediction and structure denoising and reconstruction tasks, and pre-train the crystal structure encoder (as shown in the figure neural network) on the set tasks. The trained crystal structure encoder can extract three-dimensional physical and topological features.

[0090] Through the above tasks, the system's modal base models can independently and fully learn the general basic representation without manual labeling.

[0091] Step 6: Identify entities in the current query context through an entity prediction mechanism and extract feature vectors of the identified entities. Then , , and After splicing, the splicing result is input into the Masked Multi-Cross Attention layer.

[0092] Deep feature interaction is achieved by masking multiple cross-attention layers, and a masking mechanism is used to prevent the leakage of future information during training.

[0093] Step 7: The output of the masked multi-cross attention mechanism layer is weighted and fused by average pooling, and the fusion result is fed into the generative language model (LLM is used in this invention). The generative language model uses the right-shifted output temporal mechanism to generate a response containing graph reasoning, crystal structure and material properties.

[0094] Furthermore, for special magnesium alloys, a lightweight adapter module can be added to the aforementioned network structure. The adapter parameters are trained separately using high-value data on the specific magnesium alloy, thus distinguishing the common characteristics of conventional magnesium alloys from the unique properties of special magnesium alloys. When processing conventional magnesium alloys, only the main parameters are needed, without requiring the adapter module. Therefore, by adding the adapter, high-precision professional adaptation can be achieved.

[0095] like Figure 1 and Figure 5 As shown, the diffusion model and generative language model used in this invention are trained based on federated learning. The federated learning framework includes a central server and a global model, multiple local clients and local models. The server issues tasks through federated learning and collaborates with each client to train the local model. Before aggregating into global model parameters, the server introduces a contribution evaluation mechanism to quantify the contribution of each client's data to the local parameters of the global model. Then, the global model parameters are updated to complete the overall model training.

[0096] During training, magnesium alloy crystal structures generated by the diffusion model are projected into a virtual space as initial candidates. Further, based on active learning query strategies (such as outlier detection, greedy sampling, and minimum change sampling utility functions), the candidate structures are precisely screened. High-potential candidate magnesium alloy crystal structures are then subjected to real physical experiments (such as DFT and XRD analysis) by researchers. The data of crystal structures that pass experimental verification are added to the training set of federated learning for the next round of federated learning and iteration. As this iterative process continues, the accuracy of the system in synthesizing and predicting new magnesium alloys gradually approaches the actual physical laws. Moreover, the method of this invention can respond to privacy requests and model forgetting from various clients. When a continuous or forgotten learning request (such as data withdrawal) is received from a specific client, the system triggers a forgetting mechanism to eliminate the impact of that client's data on the model. This invention provides two parallel model forgetting and retraining paths:

[0097] Path 1 (Retraining based on historical rollback): The server rolls back the global model parameters to the historical checkpoint before the addition of data for that specific user. Then, it relies on the data of other retained users to perform anti-forgetting retraining from that node to restore the model's generalization performance.

[0098] Path 2 (Local Parameter Direct Adjustment Algorithm): Instead of full retraining and historical rollback, the server calculates the Hessian matrix or influence function based on the model structure, accurately identifies and quantifies the parameter distribution containing the contribution of a specific user, and then directly performs targeted parameter subtraction and local modification on the current global model to mathematically eliminate the user's contribution.

[0099] Regardless of the path chosen, after retraining, the server continues to initiate federated learning tasks to restore and improve the prediction accuracy of the global model. This invention eliminates the need to aggregate sensitive experimental data from various institutions to a central server, achieving a complete data flow from "data processing to crystal generation to dialogue reasoning" while simultaneously meeting the privacy compliance requirements of high-value material formulations.

[0100] Specific Implementation Method Two: This implementation method describes a material property intelligent dialogue system based on knowledge graph embedding and diffusion generation mechanisms. The system includes a knowledge graph construction module, a self-supervised pre-training module, a diffusion model, and a generative language model, wherein:

[0101] The knowledge graph construction module is used to generate a knowledge graph based on the material dataset, and to construct node, edge and community indexes for the knowledge graph data to support efficient retrieval in subsequent steps;

[0102] The knowledge graph is generated using either method a or method b:

[0103] Method a: The material data in the training set is labeled in BIO format using the BERT-BiLSTM-Att-CRF model, and then the BIO-labeled dataset is retrieved using the GraphRAG retrieval tool (which includes query expansion, query decomposition, merging and pruning modules). The entities include magnesium alloy composition, crystal structure parameters and mechanical property indicators.

[0104] The recalled entities and relationships are then format-converted and aggregated. The aggregation results are then converted into adjacency lists and natural language forms to input the generator. The generator output is then converted into code snippets and syntax trees (to enhance the logic). Finally, the code snippets and syntax trees are represented as node sequences and graph embeddings. The node sequences and graph embeddings are then aggregated by the graph generator to generate a knowledge graph containing nodes and edges.

[0105] Method b: Use the DeepSeek R1 generative model to generate natural text responses with hidden entities for dialogues in the dataset, extract the sparse attention weights hidden in the natural text responses, and then use the Alignment & Parsing Module based on rules and semantic similarity to map the sparse attention weights into structured entities and relationships. Finally, import the parsed entities and relationships into GraphRAG to construct a dynamically evolving knowledge graph.

[0106] The self-supervised pre-training module is used to train the text encoder and the crystal structure encoder.

[0107] Step 1: Dataset Collection and Text Preprocessing: Collect multimodal datasets containing magnesium alloy composition, crystal structure, and mechanical properties, and perform cleaning, encoding, sequence padding, and normalization on the collected datasets;

[0108] Step 2: For text data: Set up Masked Language Modeling (MLM) and Autoregressive Language Modeling (ARLM) tasks, and pre-train the text encoder on the set tasks. The trained text encoder can extract domain semantic features.

[0109] For crystal structure data: Set up masked atom prediction and structure denoising and reconstruction tasks, and pre-train the crystal structure encoder (as shown in the figure neural network) on the set tasks. The trained crystal structure encoder can extract three-dimensional physical and topological features.

[0110] The diffusion model is a U-Net-based diffusion model, where the convolutional layers in U-Net are replaced with multi-layer graph convolutional networks (GCNs) or graph attention networks. The diffusion model takes completely noisy data and the target attributes of the current dialogue context as input to generate a crystal structure containing alloy components; specifically:

[0111] Step A: Initialize the time step The complete noise data and time-step encoding are used as input to the constructed diffusion model;

[0112] Step B: Extract node (e.g., magnesium alloy atom) features and topological edge features through multi-layer graph convolutional network (GCN) or graph attention network (GAT) in the diffusion model to complete the feature dimensionality reduction and encoding of the high-dimensional latent space;

[0113] Step C: Perform cross-attention calculation on the target attribute and the extracted node features. That is, use the target attribute as the query vector and the extracted node features as the key vector and value vector. Use the cross-attention calculation result as the updated node features to continue propagating forward. The cross-attention guides the denoising process to focus on the local atomic structures that are strongly related to the target attribute until the denoised crystal structure containing the alloy composition is output at the current time step.

[0114] By utilizing the skip-connections mechanism of U-Net, the decoder fuses deep semantic features with shallow high-resolution structural features during the denoising process, and gradually predicts the distribution of noise added at the current time step.

[0115] Step D: Determine if the condition is met. :

[0116] If satisfied, then execute step E on the crystal structure containing alloy composition output from the last time step;

[0117] If not satisfied, then let , will the The denoised crystal structure and time-step encoding after each time step are used as input to the constructed diffusion model, and the process returns to step B.

[0118] Step E: Correct the atomic positions and lattice parameters in the crystal structure output in the last time step according to the physical conservation laws and spatial topology rules of materials science, and finally output the physical crystal structure corresponding to the current query context dialogue.

[0119] The generative language model is used to generate responses that include graph reasoning, crystal structure, and material properties, specifically:

[0120] The context of the current dialogue is used as input to the text encoder, which outputs a semantic feature vector. ;

[0121] The physical crystal structure corresponding to the current query context is used as the input to the crystal structure encoder, which outputs the crystal structure feature vector. ;

[0122] The knowledge graph is retrieved based on the target attributes of the current dialogue context, and the retrieved entity nodes and relation edges are transformed into historical feature vectors of the graph. ;

[0123] Extract the feature vectors of entities in the current dialogue context. ;

[0124] right , , and The splicing process is performed, and the spliced ​​result is sequentially passed through a masking multi-cross attention mechanism layer and an average pooling operation. The processed result is then fed into a generative language model to generate a response that includes graph reasoning, crystal structure, and material properties.

[0125] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for intelligent dialogue about material properties based on knowledge graph embedding and diffusion generation mechanisms, characterized in that, The method specifically includes the following steps: Step 1: Obtain the labeled material dataset; Step 2: Construct a knowledge graph based on the acquired material dataset. After generating the knowledge graph, construct node, edge, and community indexes for the knowledge graph data. Step 3: Perform semantic decomposition on the dialogue context of the current query, extract the target attributes of magnesium alloy expected by the user, and retrieve the constructed knowledge graph based on the extracted target attributes of magnesium alloy expected by the user to retrieve the entity nodes and relation edges related to the current query. Step 4: Convert the retrieved entity nodes and relation edges into graph historical feature vectors. The semantic feature vector of the current query dialogue context is extracted using a text encoder. ; Construct a diffusion model based on U-Net, and replace the convolutional layers in U-Net with a multi-layer graph convolutional network and a graph attention network; The user's desired magnesium alloy target properties and noise are used as inputs to the diffusion model, which generates a crystal structure containing the alloy composition. Step 5: Use a crystal structure encoder to extract features from the physical crystal structure generated in Step 4 to obtain the crystal structure feature vector. ; Step 6: Identify entities in the current query context through an entity prediction mechanism and extract feature vectors of the identified entities. Then , , and After splicing, the splicing result is input into the masking multi-cross attention mechanism layer; Step 7: The output of the masked multi-cross attention mechanism layer is weighted and fused by average pooling operation. The fusion result is fed into the generative language model. The generative language model uses the temporal mechanism of right-shifting output to generate a response containing graph reasoning, crystal structure and material properties.

2. The intelligent dialogue method for material properties based on knowledge graph embedding and diffusion generation mechanism according to claim 1, characterized in that, The knowledge graph is constructed in the following way: The BERT-BiLSTM-Att-CRF model was used to label the material data in the training set in BIO format. Then, the GraphRAG retrieval tool was used to recall entities and relationships in the BIO-labeled dataset. The entities include magnesium alloy composition, crystal structure parameters and mechanical property indicators. The recalled entities and relationships are then format-converted and aggregated. The aggregation results are then converted into adjacency lists and natural language forms to input the generator. The generator output is then converted into code snippets and syntax trees. Finally, the code snippets and syntax trees are represented as node sequences and graph embeddings. The node sequences and graph embeddings are then aggregated by the graph generator to generate a knowledge graph containing nodes and edges.

3. The intelligent dialogue method for material properties based on knowledge graph embedding and diffusion generation mechanism according to claim 1, characterized in that, The knowledge graph is constructed in the following way: We use the DeepSeek R1 generative model to generate natural text responses with hidden entities for dialogues in the dataset, and extract the sparse attention weights hidden in the natural text responses. Then, we use the alignment parsing module based on rules and semantic similarity to map the sparse attention weights into structured entities and relationships. Finally, we import the parsed entities and relationships into GraphRAG to construct a dynamically evolving knowledge graph.

4. The intelligent dialogue method for material properties based on knowledge graph embedding and diffusion generation mechanism according to claim 1, characterized in that, The process involves using the user's desired magnesium alloy target properties and noise as input to the diffusion model, and generating a crystal structure containing the alloy composition through the diffusion model. Specifically: Step 4.1 Initialize the time step The complete noise data and time-step encoding are used as input to the constructed diffusion model; Step 4.2: Extract node features and topological edge features using a multi-layer graph convolutional network or graph attention network within the diffusion model; Step 4.3: Perform cross-attention calculation on the target attribute and the extracted node features. That is, use the target attribute as the query vector, the extracted node features as the key vector and value vector, and use the cross-attention calculation result as the updated node features to continue propagating forward until the denoised crystal structure containing the alloy composition is output at the current time step. Step 44: Determine if the condition is met. : If satisfied, then execute steps four and five on the crystal structure output at the last time step; If not satisfied, then let , will the The denoised crystal structure and time-step encoding after each time step are used as input to the constructed diffusion model, and the process returns to step four two. Steps four and five: Based on the physical conservation laws and spatial topology rules of materials science, the types and positions of atoms and the lattice parameters in the crystal structure output in the last time step are corrected and discretized, and finally the physical crystal structure corresponding to the current query context dialogue is output.

5. The intelligent dialogue method for material properties based on knowledge graph embedding and diffusion generation mechanism according to claim 4, characterized in that, The training process for the text encoder and the crystal structure encoder is as follows: Step 1: Dataset Collection and Text Preprocessing: Collect multimodal datasets containing magnesium alloy composition, crystal structure, and mechanical properties, and perform cleaning, encoding, sequence padding, and normalization on the collected datasets; Step 2, Self-supervised pre-training: For text data: Set up masked language modeling and autoregressive language modeling tasks, and pre-train the text encoder on the set tasks; For crystal structure data: Set up mask atom prediction and structure denoising and reconstruction tasks, and pre-train the crystal structure encoder on the set tasks.

6. A material property intelligent dialogue system based on knowledge graph embedding and diffusion generation mechanism, characterized in that, The system includes a knowledge graph construction module, a self-supervised pre-training module, a diffusion model, and a generative language model, wherein: The knowledge graph construction module is used to generate a knowledge graph based on the materials dataset. The self-supervised pre-training module is used to train the text encoder and the crystal structure encoder. The diffusion model is a U-Net-based diffusion model, in which the convolutional layers in U-Net are replaced by multi-layer graph convolutional networks or graph attention networks. The diffusion model takes the complete noisy data and the target attributes of the current dialogue context as input to generate a crystal structure containing alloy components. The generative language model is used to generate responses that include graph reasoning, crystal structure, and material properties.

7. The intelligent dialogue system for material properties based on knowledge graph embedding and diffusion generation mechanism according to claim 6, characterized in that, The knowledge graph construction method is as follows: The BERT-BiLSTM-Att-CRF model was used to label the material data in the training set in BIO format. Then, the GraphRAG retrieval tool was used to recall entities and relationships in the BIO-labeled dataset. The entities include magnesium alloy composition, crystal structure parameters and mechanical property indicators. The recalled entities and relationships are then format-converted and aggregated. The aggregation results are then converted into adjacency lists and natural language forms to input the generator. The generator output is then converted into code snippets and syntax trees. Finally, the code snippets and syntax trees are represented as node sequences and graph embeddings. The node sequences and graph embeddings are then aggregated by the graph generator to generate a knowledge graph containing nodes and edges.

8. The intelligent dialogue system for material properties based on knowledge graph embedding and diffusion generation mechanism according to claim 6, characterized in that, The knowledge graph construction method is as follows: We use the DeepSeek R1 generative model to generate natural text responses with hidden entities for dialogues in the dataset, and extract the sparse attention weights hidden in the natural text responses. Then, we use the alignment parsing module based on rules and semantic similarity to map the sparse attention weights into structured entities and relationships. Finally, we import the parsed entities and relationships into GraphRAG to construct a dynamically evolving knowledge graph.

9. The intelligent dialogue system for material properties based on knowledge graph embedding and diffusion generation mechanism according to claim 6, characterized in that, The working process of the self-supervised pre-training module is as follows: Step 1: Dataset Collection and Text Preprocessing: Collect multimodal datasets containing magnesium alloy composition, crystal structure, and mechanical properties, and perform cleaning, encoding, sequence padding, and normalization on the collected datasets; Step 2: For text data: Set up masked language modeling and autoregressive language modeling tasks, and pre-train the text encoder on the set tasks; For crystal structure data: Set up mask atom prediction and structure denoising and reconstruction tasks, and pre-train the crystal structure encoder on the set tasks.

10. The intelligent dialogue system for material properties based on knowledge graph embedding and diffusion generation mechanism according to claim 9, characterized in that, The working process of the diffusion model is as follows: Step A: Initialize the time step The complete noise data and time-step encoding are used as input to the constructed diffusion model; Step B: Extract node features and topological edge features through a multi-layer graph convolutional network or graph attention network within the diffusion model; Step C: Perform cross-attention calculation on the target attribute and the extracted node features. That is, use the target attribute as the query vector, the extracted node features as the key vector and value vector, and use the cross-attention calculation result as the updated node features to continue propagating forward until the denoised crystal structure containing alloy composition is output at the current time step. Step D: Determine if the condition is met. : If satisfied, then execute step E on the crystal structure containing alloy composition output from the last time step; If not satisfied, then let , will the The denoised crystal structure and time-step encoding after each time step are used as input to the constructed diffusion model, and the process returns to step B. Step E: Based on the physical conservation laws and spatial topology rules of materials science, the atomic types, atomic positions, and lattice parameters in the crystal structure output at the last time step are corrected and discretized, and finally the physical crystal structure corresponding to the current query context dialogue is output.

Images